Maximized Setup and Other Questions

[mbn451]

Hey so I am fairly new to the site and crunching, and I have a few basic, newbie questions.

1.) Say I had 2-3k to spend trying to find a world record prime, what would be the best way to spend it? Would it be better to just build a really strong single machine, or to try building a cluster that I could add to if/when I wanted to spend more resources towards the effort?

2.) If building a cluster is the best option, what is a good starting point? I am a biologist by profession, so my computer skills are only hobbyist level-but I really want to learn more!

3.) I am currently running Prime95 on a desktop with an Intel core i5 750, and it seems to be underperforming. Interestingly each worker seems to be slower than the next, with worker 1 being 60% done and worker 4 being only 40% done with a first time LL. The exponents are slightly higher on each core, but I don't see a similar trend with my quad core iMac that is blazing away at LL.

4.) Which factoring method (P-1 or TF) would yield the most work units per day? Would they theoretically be the same?

5.) Any suggested way to "clean up" a computer so that it runs faster? I have of course gone into task manager and ended processes that seem non-essential (a slightly harder task than it sounds considering confusing process names) but is there anything else I can do?

If anybody has any insights I would be happy to hear them, thanks!

[Dubslow]

1-2) Finding the system that has the best price/performance ratio and then getting lots of that is the best way to build a cluster. An i5-3550K is definitely up there in terms of performance for a good price, but there might be other processors that are cheaper with a better price/performance (but keep in mind that each processor usually needs its own motherboard/case/etc, so there is a certain amount of fixed overhead per processor...)

4) They would both yield the same credit from a CPU, however trial factoring can be run much much faster on GPUs than CPUs, so it's generally best if CPUs don't do TF and instead do LL or P-1. Additionally, P-1 is in seriously short supply at the moment -- the problem of course is that it requires a decent chunk of RAM (1 GiB would be sufficient).

5) Not really much other than what you have just described.

[Mr. P-1]

Originally Posted by mbn451

Hey so I am fairly new to the site and crunching, and I have a few basic, newbie questions.

1.) Say I had 2-3k to spend trying to find a world record prime, what would be the best way to spend it? Would it be better to just build a really strong single machine, or to try building a cluster that I could add to if/when I wanted to spend more resources towards the effort?

If by "cluster" you mean a bunch of nodes with high-speed communications between them, then the latter is not necessary. A farm of independent processors would do just as well.

To get the best bang for your buck, you need to focus your expenditure on the processors and the memory, bearing in mind also the cost of electricity over the lifetime of the hardware. The motherboard is important to the extent that it doesn't introduce bottlenecks in memory access, but otherwise you can do without the bells and whistles a high-end MoBo will give you.

2.) If building a cluster is the best option, what is a good starting point? I am a biologist by profession, so my computer skills are only hobbyist level-but I really want to learn more!

My budget for my last build, just over a month ago, was £500, and I wanted to play around with overclocking, so a core i5 3570K system was the obvious choice for me. If you're not intending to overclock, you could dispense with the K suffix and the fancy cooler and get basically the same system for under £400. On your budget* you could buy five or more of these. The corresponding i7 gives an extra 100MHz and hyperthreading, which isn't a lot of benefit to GIMPS, so didn't justify the extra cost.

*It used to be the case that you could buy about as much hardware for $1 in the US as £1 in the UK. I don't know if this still applies.

3.) I am currently running Prime95 on a desktop with an Intel core i5 750, and it seems to be underperforming. Interestingly each worker seems to be slower than the next, with worker 1 being 60% done and worker 4 being only 40% done with a first time LL. The exponents are slightly higher on each core, but I don't see a similar trend with my quad core iMac that is blazing away at LL.

Check your FFT lenghs. If they are the same, then running time ought to be proportional to the exponent. If the higher exponent uses a larger FFT then this will take significantly more time.

4.) Which factoring method (P-1 or TF) would yield the most work units per day? Would they theoretically be the same?

Time to complete a work-unit depends upon the exponent and the amount of Prior TF. The amount of credit you get for completing one ought to be roughly proportional to the time you spend on it (unless George has rejigged the TF credit recently. GPUs have rendered CPUs obsolete for this type of work.)

GIMPS has a shortage of dedicated P-1 capacity. IIRC you need at least 300MB of RAM. 1GB per core doing stage 2 is ample.

(Bear in mind that you cannot find record-breaking primes by doing factoring work.)

5.) Any suggested way to "clean up" a computer so that it runs faster? I have of course gone into task manager and ended processes that seem non-essential (a slightly harder task than it sounds considering confusing process names) but is there anything else I can do?

Install Linux.

[Mr. P-1]

Originally Posted by Dubslow

An i5-3550K is definitely up there in terms of performance for a good price, but there might be other processors that are cheaper with a better price/performance (but keep in mind that each processor usually needs its own motherboard/case/etc, so there is a certain amount of fixed overhead per processor...)

Consider also the cost of electricity over the lifetime of the equipment. This tips the balance in favor of fewer, faster machines.

[Dubslow]

Originally Posted by Mr. P-1

Consider also the cost of electricity over the lifetime of the equipment. This tips the balance in favor of fewer, faster machines.

No, usually it tips the balance in favor of copious but slower machines -- lots of little ARM processors would be most efficient (but can't run Prime95). Power consumption scales exponentially with speed -- the lower down the curve you are, typically the better. (That said, an AMD consumes a lot more power per performance than an Intel, at the moment anyways, hence the 3550 3570 recommendation we both made.)

[Mr. P-1]

Originally Posted by Dubslow

No, usually it tips the balance in favor of copious but slower machines -- lots of little ARM processors would be most efficient (but can't run Prime95). Power consumption scales exponentially with speed -- the lower down the curve you are, typically the better. (That said, an AMD consumes a lot more power per performance than an Intel, at the moment anyways, hence the 3550 3570 recommendation we both made.)

I was assuming that we were only talking about machines which can run Prime95/mprime. My thinking was that in addition to the power consumed by the processor, there would also be an essentially fixed overhead (per machine) associated with the remaining components. More machines = more overhead.

[Dubslow]

Originally Posted by Mr. P-1

I was assuming that we were only talking about machines which can run Prime95/mprime. My thinking was that in addition to the power consumed by the processor, there would also be an essentially fixed overhead (per machine) associated with the remaining components. More machines = more overhead.

For a compute-only cluster, the overhead of a mobo and HDD/SDD is a few tens of watts at most -- often less, and so doesn't play too much of a role. Even if we limit ourselves to x86, you'd be better off with a whole bunch of 15 W processors (30W box total) -- but they cost similar, so that's only better from a power-only perspective, not from a capital/performance perspective.

Either way though, we both know what we're talking about since we both recommended the same proc

[swl551]

I totally agree i5 is more cost effective (cost, heat, power, performance) and scaling out, not up is the better choice.

Buying 2 mid-range i5 pcs can cost less than a tricked-out* i7. Total power consumption would be a bit more, but overall productivity would be about double a single i7.


* is referring to the "ultra-gaming" type setup with mobos costing > $300 etc.


Both my i5s run on mobo < $80.00 and they kick my i7 to the ground.

[lycorn]

A very important point, as far as performance goes, is avoiding memory bottlenecks. This is particularly important when using AVX capable CPUs running a 64-bit client. A constrained memory access may ruin the CPU performance, when running several tests in parallel. As an example, I´ve recently tested a PC with an i5-3450, and the per iteration time, for the same FFT size on a LL test, was 0.012s with one worker running and 0.024s with 4 workers running. The mobo was nothing to write home about, nor was the memory. Therefore, I recommend investing in memory bandwidth!

[Uncwilly]

Originally Posted by Dubslow

For a compute-only cluster, the overhead of a mobo and HDD/SDD is a few tens of watts at most -- often less, and so doesn't play too much of a role.

If you network boot them (or use a USB drive to boot it) you don't have to buy a HDD or run it. Multiple by number of machines...

[mbn451]

Thanks for all the feedback! Looking at the Benchmarks page, why is a 3570k recommended over a 2500k? It seems as if the 2500k has better times on FFT lengths, and is currently the same price as the 3570k on Newegg. What am I missing here?